Data Centers. Tom Anderson

Transcription

1 Data Centers Tom Anderson

2 Transport Clarification RPC messages can be arbitrary size Ex: ok to send a tree or a hash table Can require more than one packet sent/received We assume messages can be dropped, duplicated, reordered It is packets that are dropped, duplicated, reordered Most RPCs are single packets TCP addresses some of these issues But we ll assume the general case

3 RPC Semantics At least once (NFS, DNS, lab 1b) true: executed at least once false: maybe executed, maybe multiple times At most once (lab 1c) true: executed once false: maybe executed, but never more than once Exactly once true: executed once false: never returns false

4 When does at-least-once work? If no side effects read-only operations (or idempotent ops) Example: MapReduce Example: NFS readfileblock (vs. Posix file read) writefileblock (but not: delete file)

5 At Most Once Client includes unique ID (UID) with each request use same UID for re-send Server RPC code detects duplicate requests return previous reply instead of re-running handler if seen[uid] { r = old[uid] } else { r = handler() old[uid] = r seen[uid] = true }

6 Some At-Most-Once Issues How do we ensure UID is unique? Big random number? Combine unique client ID (IP address?) with seq #? What if client crashes and restarts? Can it reuse the same UID? In labs, nodes never restart Equivalent to: every node gets new ID on start Client address + seq # is unique Seq # alone is NOT unique

7 When Can Server Discard Old RPCs? Option 1: Never? Option 2: unique client IDs per-client RPC sequence numbers client includes "seen all replies <= X" with every RPC Option 3: one client RPC at a time arrival of seq+1 allows server to discard all <= seq Labs use Option 3

8 What if Server Crashes? If at-most-once list of recent RPC results is stored in memory, server will forget and accept duplicate requests when it reboots Does server need to write the recent RPC results to disk? If replicated, does replica also need to store recent RPC results? In Labs, server gets new address on restart Client messages aren t delivered to restarted server (discard if sent to wrong version of server)

9 A Data Center

10 Inside a Data Center

11 Data center 10k - 100k servers: 250k 10M cores 1-100PB of DRAM 100PB - 10EB storage 1-10 Pbps bandwidth (>> Internet) MW power - 1-2% of global energy consumption 100s of millions of dollars

12 Servers Limits driven by the power consumption 1-4 multicore sockets cores/socket (150W each) 100s GB 1 TB of DRAM ( W) 40Gbps link to network

13 19 wide 1.75 tall (1u) (defined in 1922!) servers/rack network at top Servers in racks

14 Racks in rows

15 Rows in hot/cold pairs

16 Hot/cold pairs in data centers

17 Amazon, in the US: - Northern Virginia - Ohio - Oregon - Northern California Where is the cloud? Why those locations?

18 MTTF/MTTR Mean Time to Failure/Mean Time to Repair Disk failures (not reboots) per year ~ 2-4% At data center scale, that s about 2/hour. It takes 10 hours to restore a 10TB disk Server crashes 1/month * 30 seconds to reboot => 5 mins/year 100K+ servers

19 Typical Year in a Data Center (2008) ~0.5 overheating (power down most machines in <5 mins, ~1-2 days to recover) ~1 PDU failure (~ machines suddenly disappear, ~6 hours to come back) ~1 rack-move (plenty of warning, ~ machines powered down, ~6 hours) ~1 network rewiring (rolling ~5% of machines down over 2-day span) ~20 rack failures (40-80 machines instantly disappear, 1-6 hours to get back)

20 Typical Year in a Data Center (2008) ~5 racks go wonky (40-80 machines see 50% packetloss) ~8 network maintenances (4 might cause ~30-minute random connectivity losses) ~12 router reloads (takes out DNS and external vips for a couple minutes) ~3 router failures (have to immediately pull traffic for an hour) ~dozens of minor 30-second blips for dns ~1000 individual machine failures ~thousands of hard drive failures slow disks, bad memory, misconfigured machines, flaky machines, etc

21 This Week: Primary Backup How do we build systems that survive a single node failure? State replication: run two copies of server primary, backup different racks, different PDUs, different DCs! If primary fails, backup takes over How do we know primary failed vs. is slow?

22 Data Center Networks Every server wired to a ToR (top of rack) ToR s in neighboring aisles wired to an aggregation Agg. es wired to core es

23 Early data center networks 3 layers of es - Edge (ToR) - Aggregation - Core

24 Early data center networks 3 layers of es - Edge (ToR) - Aggregation - Core Optical Electrical

25 Early data center limitations Cost - Core, aggregation routers = high capacity, low volume - Expensive! Fault-tolerance - Failure of a single core or aggregation router = large bandwidth loss Bisection bandwidth limited by capacity of largest available router - Google s DC traffic doubles every year!

26 Clos networks How can I replace a big by many small es? Big

27 Clos networks How can I replace a big by many small es? Big

28 Clos Networks What about bigger es?

29

30

31 Multi-rooted tree Every pair of nodes has many paths Fault tolerant! But how do we pick a path?

32 Multipath routing Lots of bandwidth, split across many paths ECMP: hash on packet header to determine route - (5 tuple): Source IP, port, destination IP, port, prot. - Packets from client server usually take same route On or link failure, ECMP sends subsequent packets along a different route => Out of order packets!

33 Data Center Network Trends RT latency across data center ~ 10 usec 40 Gbps links common, 100 Gbps on the way 1KB packet every 80ns on a 100Gbps link Direct delivery into the on-chip cache (DDIO) Upper levels of tree are (expensive) optical links Thin tree to reduce costs Within rack > within aisle > within DC > cross DC Latency and bandwidth: keep communication local

34 Local Storage Magnetic disks for long term storage High latency (10ms), low bandwidth (250MB/s) Compressed and replicated for cost, resilience Solid state storage for persistence, cache layer 50us block access, multi-gb/s bandwidth Emerging NVM Low energy DRAM replacement Sub-microsecond persistence

35 Day in the Life of a Web Request User types google.com, what happens DNS = Domain Name System Global database of name->ip address DNS query to root name server Root name server is replicated (600+) Hard coded IP multicast addresses Updates happen async in background (rare) Root returns: IP address of google s name server Cached on client (for a configurable period) Can be out of date

36 Name Server to DC DNS returns google s name server Google name server returns local name server Name server in DC close to user s IP address Local name server returns IP of web front end Also in local data center With short timeout, reroute if front end failure Browser sends HTTP request to front end

37 Resilient Scalable Web Front End IP address of front end is an array of machines Packets first go through load balancer Actually, an array of load balancers (all the same) Load balancer hashes 3 or 5 tuple -> front end Source IP, port #, Destination IP, port #, protocol This way, all traffic from user goes to same server When front end fails, load balancer redirects incoming packets elsewhere

38 Challenge How do we build a hash function that s stable? That reroutes only the packets for the failed node, leaving the packets to other nodes alone.

39 Web Front End -> Cache, Storage Layer Key-value store Sharded across many cache and storage nodes Hash(key) -> cache, storage node If cache node fails, rehash to new cache node If storage node fails,???